This is drilling down on the issue of why multi-stage lookup tables of new std.uni ahve decent speed in LDC and painstakingly slow in DMD.
Observe that the following snippet (inlined opIndex of a Trie) does 2 remarkably stupid things:
a) See that read/write of arg_0 on stack, instead of direct "mov edx, ebx"
b) 2 push eax at the begining of function ... and "add esp, 8" at end - WAT? Note that eax is never written to until the very end (there is simply no need to save it).
public _D3std3uni146__T4TrieTS3std3uni19__T9BitPackedTbVi1Z9BitPackedTwVk1114112TS3std3uni21__T9sliceBi39F8D5E3EE00191D27B7780CD5A2FFED
_D3std3uni146__T4TrieTS3std3uni19__T9BitPackedTbVi1Z9BitPackedTwVk1114112TS3std3uni21__T9sliceBi39F8D5E3EE00191D27B7780CD5A2FFED proc near
; CODE XREF: _D9trie_test4mainFAAyaZv17__foreachbody6846MFNfKwZi+Ap
var_8 = dword ptr -8
arg_0 = dword ptr 4
push eax
mov ecx, [esp+4+arg_0]
shr ecx, 8
push eax
and ecx, 1FFFh
mov edx, [eax+14h]
push ebx
mov bx, [edx+ecx*2]
push esi
mov esi, [esp+10h+arg_0]
and esi, 0FFh
push edi
mov edi, [eax+4]
lea ecx, [edx+edi*4]
and ebx, 0FFFFh
shl ebx, 8
add ebx, esi
mov [esp+14h+var_8], ebx
mov edx, [esp+14h+var_8]
shr ebx, 5
and edx, 1Fh
bt [ecx+ebx*4], edx
sbb eax, eax
neg eax
pop edi
pop esi
pop ebx
add esp, 8
retn 4
_D3std3uni146__T4TrieTS3std3uni19__T9BitPackedTbVi1Z9BitPackedTwVk1114112TS3std3uni21__T9sliceBi39F8D5E3EE00191D27B7780CD5A2FFED endp
And the D code where the above object code can be seen:
//Command line: dmd -O -release -inline -noboundscheck
import std.uni, std.stdio;
//match bits that std.regex had before
alias codepointSetTrie!(13, 8) makeTrie;
void main(string argv[])
{
auto tr = makeTrie(unicode.Alphabetic);
int count;
foreach(arg; argv)
foreach(dchar ch; arg)
if(tr[ch])
count++;
writeln(count);
}
Comment #1 by dmitry.olsh — 2013-08-30T13:58:54Z
> See that read/write of arg_0 on stack ...
Should be var_8, of course.
Comment #2 by bugzilla — 2015-09-23T07:19:30Z
The PUSH EAX is there to align the stack to 16 bytes.
Comment #3 by robert.schadek — 2024-12-13T18:11:05Z